The Evolution of Indoor Positioning Systems
The field of human activity recognition has evolved significantly, driven largely by advancements in Internet of Things (IoT) device technology, particularly in personal devices. This study investigates the use of ultra-wideband (UWB) technology for tracking inhabitant paths in home environments using deep learning models.
UWB technology estimates user locations via time-of-flight and time-difference-of-arrival methods, which are significantly affected by the presence of walls and obstacles in real environments, reducing their precision. To address these challenges, we propose a fingerprinting-based approach utilizing received signal strength indicator (RSSI) data collected from inhabitants in two flats (60 m² and 100 m²) while performing daily activities.
Comparing Modeling Approaches
We compare the performance of convolutional neural network (CNN), long short-term memory (LSTM), and hybrid CNN+LSTM models, as well as the use of Bluetooth technology. Additionally, we evaluate the impact of the type and duration of the temporal window (future, past, or a combination of both).
Our results demonstrate a mean absolute error close to 50 cm, highlighting the superiority of the hybrid model in providing accurate location estimates, thus facilitating its application in daily human activity recognition in residential settings.
Overcoming Limitations of Traditional Positioning Methods
Traditional indoor positioning methods often struggle with signal obstruction, environmental variability, and high infrastructure costs. UWB technology offers higher precision than Bluetooth Low Energy (BLE), but it also faces significant challenges that affect its performance. Signal obstruction is a primary concern, as UWB signals are susceptible to attenuation and multipath effects caused by physical obstructions such as walls and furniture, leading to reduced accuracy in location estimation.
To address these challenges and improve precision, we focus on radio-based technologies, specifically UWB and BLE. Our approach involves deploying anchors throughout the facility to receive signals from tags attached to the entities. Initially, the user maps their location within the environment, and these labels are collected and correlated with the signals received by the anchors. These data are sent to a server for processing, where the model learns from the labelled data. Once trained, the system can accurately calculate the location in real-time of the entity using only RSSI signals, reducing costs and facilitating deployment in real environments.
Fingerprinting Techniques for Improved Accuracy
Fingerprinting, a common technique used in location systems, involves creating a map of signal strengths at various locations and using this map to estimate positions. Although effective, this method requires a substantial amount of labelled data, and the performance of BLE can be inconsistent due to signal fluctuations.
To address the problem of non-line-of-sight (NLOS) conditions and reduce the impact of the number of anchors required, the fingerprinting method plays a crucial role. By correlating the real-world coordinates with the RSSI values received from each beacon, fingerprinting can provide accurate location estimates even in complex environments. This method mitigates the limitations of UWB in obstructed settings by leveraging the consistent patterns of signal strengths mapped during the fingerprinting process.
Leveraging Deep Learning Models
To enhance the accuracy of indoor positioning systems, incorporating advanced deep learning architectures, such as CNNs and LSTMs, with fingerprinting techniques is highly beneficial. CNNs are proficient in detecting spatial patterns in RSSI data, while LSTMs are adept at capturing temporal dependencies and dynamics in signal sequences. Using these deep learning models improves location estimation accuracy, better adapts to the variability of indoor environments, and reduces the reliance on a dense network of anchors.
The combination of CNN and LSTM allows for the creation of more sophisticated and adaptive models that can provide precise real-time positioning in complex indoor settings. This can be corroborated in a variety of works in different areas, highlighting the potential of combining spatial and temporal data, handling data intricacies, and incorporating attention mechanisms to improve prediction accuracy.
Case Study and Experimental Validation
To evaluate the effectiveness of our approach, we conducted a case study in two flats (60 m² and 100 m²) with different device configurations and sensor deployments. The data used in our model includes the location data tagged on the map by the user, serving as the ground truth, as well as the RSSI data collected from UWB anchors and BLE beacons.
Our experimental results demonstrate a mean absolute error close to 50 cm, highlighting the superiority of the hybrid CNN+LSTM model in providing accurate location estimates. Additionally, we analyzed the impact of the type and duration of the temporal window, finding that increasing the window size reduces the error, although with diminishing returns.
Conclusion and Future Directions
This study showcases the potential of UWB technology and deep learning models for tracking daily paths in home contexts. By leveraging RSSI fingerprinting and advanced deep learning architectures, we have developed a robust and cost-effective indoor positioning system that can accurately estimate the location of inhabitants in residential settings.
However, it is necessary to test the system in multi-occupancy environments and different architectural layouts to ensure its robustness and address potential challenges. In the future, our goal is to deploy this architecture in a real environment, operational 24/7, to gather comprehensive data and further refine our models. We will also explore the integration of ambient binary sensors to achieve better human activity recognition in homes with multiple inhabitants.
These advancements will contribute to the development of more efficient and user-friendly indoor positioning systems, significantly improving the smart home experience and paving the way for enhanced human activity recognition in complex residential environments.
